AI Demand Tightens Chip Supply Chain; Google TPU Pays for Itself in One Year; China Explores EUV-Free 3nm Route
nashnova research
Google Cloud CEO reveals custom TPU servers recoup their investment in under a year, fueling the Big Tech custom-chip arms race; meanwhile, semiconductor equipment lead times have doubled and China's CAS has disclosed an early EUV-free 3nm route — deep supply-chain constraints are now setting the real pace of AI compute expansion.
Google's TPU pays back in under a year — what does the math say?
Google Cloud CEO Thomas Kurian says overall AI server payback is under two years; servers using in-house TPUs recoup in under one year.
This means → the economics of custom silicon have cleared the bar: the faster the payback, the stronger the case to reinvest, and the custom-chip flywheel accelerates.
TPU orders keep climbing. Applications now span finance, biopharma, and high-performance computing — in plain terms = TPUs are no longer a lab experiment but a revenue-generating production tool.
This reflects a broader shift: Big Tech is proving that building your own chips beats buying general-purpose ones. Broadcom and MediaTek, as ASIC partners — firms that help design and produce custom chips — stand to benefit directly.
The 2027 TPU packaging plan — why the uncertainty?
Google's 2027 TPU project, codenamed "Humufish," was set to use Intel's EMIB-T packaging with MediaTek as co-developer.
Market sources now say substrate yield issues have prompted Google to list TSMC's CoWoS — an advanced packaging method that tiles multiple chips onto one substrate — as a backup. Substrate supplier Unimicron had targeted just 50% yield.
Multiple sources still expect EMIB-T to remain the primary plan: there is time before 2027 mass production, and running dual packaging paths would add cost and complexity.
Put simply = Google hasn't switched plans, but it has a Plan B. The real question is whether substrate yields catch up before production begins.
Equipment lead times have doubled — where is the bottleneck?
Lead times for critical semiconductor equipment components have stretched sharply: a four-month delivery has become ten months; some Japanese parts now take 40 months; local component lead times are up roughly 50%.
The backdrop: TSMC, ASML, and Applied Materials are all expanding simultaneously. SEMI forecasts wafer-fab and DRAM equipment sales will hit all-time highs.
This means → no matter how much capex flows in, actual capacity ramp stays gated by deep supply-chain constraints — the money is there, but the machine parts are not.
This reflects a pivotal shift: the real bottleneck in AI compute expansion is moving from "who writes the check" to "who can deliver the hardware."
China's EUV-free 3nm route — how far along is it?
Ye Tianchun, chief engineer at the Chinese Academy of Sciences' Institute of Microelectronics, disclosed an early-stage sub-3nm device path built without EUV lithography — the most advanced chipmaking light source — using DUV lithography (the previous generation) combined with GAA transistor architecture (gate-all-around — a structure that wraps the gate around the transistor channel on all sides for better control).
After process optimization, the key Ion/Ioff ratio exceeded 5×10⁴, reaching 9.7×10⁴ and 7.6×10⁴ respectively.
The institute itself classified this as technology validation, not production-ready, citing major integration and yield challenges ahead.
In plain terms = the lab has shown the path is walkable, but the distance from lab to factory remains vast. Its significance: under real EUV restrictions, China now has an early technical reference point.
SanDisk is hiring aggressively in South Korea — what does that signal?
SanDisk is running an unusually large recruiting push in South Korea, spanning NAND architecture, VLSI design, and mask design, with requirements for HBM, LPDDR, and GDDR experience.
This marks a sharp departure from SanDisk's historically sales-and-marketing-focused Korean operations.
This means → SanDisk is extending its R&D footprint into Samsung's and SK Hynix's talent heartland, competing directly for AI-memory engineers.
Taiwan's Physical AI opportunity — what does it take to capture it?
At a Deloitte Taiwan seminar in Taichung, speakers from ITRI, the Taiwan AI Academy, NVIDIA, and Techman Robot emphasized that Taiwan's opportunity in Physical AI — the domain where AI controls robots, automation, and the physical world — depends on systems-integration capability, not single-component supply.
Participants called for building verifiable, scalable industrial use cases.
In plain terms = Taiwan's edge is not in making any one part, but in assembling all the parts into a working system — that is the truly scarce capability in the Physical AI race.
市场有风险，内容仅供研究参考，不构成投资建议。